Semiconductor devices are specified (spec'ed) to operate within certain parameters, such as a maximum power draw and a maximum clock frequency. While semiconductor manufacturing processes are very precise, process variations do occur. Although the manufacturing process may target a typical device, sometimes process variations produce slower chips or faster chips. As device sizes shrink, larger relative variations may occur.
Chips may be tested to determine their power draw and speed, and these chips may be sorted into slow-chip bins, fast-chip bins, and typical-chip bins. The faster chips may be sold as faster speed grades, while the slower chips may be sold for slower speed grades. Unfortunately, such process skews are not always reproducible or planned but may occur randomly, making for logistical difficulties. Therefore all process skews are often lumped together. The slowest expected process skews determine the specified speed of the device, while the fastest expected process skews determine the specified maximum power dissipation.
FIG. 1 is a graph showing how process variations affect device specifications. The slowest process skew (SS) has the lowest power and the lowest performance or speed. A typical process (TT) has a better power and performance product. The fastest process skew (FF) has the highest performance and speed, but also consumes the most power.
All three process skews—slow, typical, and fast, share the same device specifications when no grade sorting is performed. Devices produced with the slowest process determine the speed specs such as the maximum clock frequency, or the minimum clock-to-output delay times. However, the fast devices consume more power than do the slower devices, so power specs are determined by devices manufactured by the fast process skews. The power-supply voltage VDD is usually fixed.
The performance and power specs are determined by the worst-case devices over the expected process skews. Slow devices set the speed specs and fast devices set the power specs. This is not optimal, since fast devices are spec'ed slower than they can actually operate, and slow devices actually draw less power than spec'ed.
Specialized sensors may be added to chips to facilitate at-speed testing. Dummy bit lines have been added to RAM arrays to adjust bit-line sensing circuits. An oscillator or a canary circuit may be added to track process variations. However, the actual circuit may be much more complex than an oscillator, resulting in tracking errors. For logic chips, a dummy path and an on-chip timing sensor may be added. The timing sensor can report its results to a tester or even to an on-chip controller that can adjust operating conditions, such as to slow down or stop a clock to reduce power consumption.
The actual critical paths in a logic circuit are the first to fail as the applied clock speed is increased. The actual critical path will have different characteristics than a dummy load to a sensor. Cross talk from neighboring nodes will differ even if gates, capacitances, and wiring traces are exactly mimicked. Extensive corner analysis may be needed to set sufficiently large margins to account for the differences between dummy paths and actual critical paths.
While such on-chip dummy paths and sensors are useful, it is desired to measure the actual critical paths rather than measure a dummy path. It is desired to add a timing sensor to an actual critical path on a chip so that the timing sensor is measuring the delay of the same physical path that carries functional data during operation of the chip.
What is desired is a critical-path timing sensor that measures delays on the same physical components that are used in functional critical paths on a chip. It is desirable to use the same gates, wires, and loads of a functional critical path. During a test mode the critical path is exercised by a test pattern to measure the timing delays of the actual functional critical path. During normal modes of operation, the functional critical path carries normal data or signals in the chip.